acceptodds
Under review as a conference paper at ICLR 2027

Blip2Mol: Phenotype–Target Fusion for Molecular Generation via Multimodal Latent Diffusion

Abstract

De novo molecular design is commonly driven either by protein structure, which favors target affinity, or by phenotypic readouts, which better reflect system-level biological activity. However, existing methods rarely integrate both modalities in a unified generative framework, limiting their ability to produce molecules that are simultaneously bioactive and target-compatible. We propose Blip2Mol, a multimodal molecular generation framework that bridges chemical-induced transcriptomic responses, protein binding pockets, and molecular structures through a shared latent space. Blip2Mol first pretrains a molecular variational autoencoder to obtain a continuous chemical representation. It then aligns transcriptome and pocket embeddings to the molecular latent space via BLIP-style dual cross-modal alignment with contrastive and matching objectives, enabling semantic consistency across phenotype, target, and molecule modalities. Next, we develop a conditional latent diffusion model with a learnable gated fusion mechanism to generate molecules jointly conditioned on transcriptomic response and pocket context. Experiments on 10 cancer-related targets show that Blip2Mol consistently outperforms representative phenotype-guided and docking-based baselines, achieving stronger docking affinity while maintaining superior drug-likeness, synthetic accessibility, validity, uniqueness, and novelty. Ablation studies further confirm the effectiveness of multimodal conditioning and the critical role of the diffusion module. These results demonstrate the promise of our unified model for phenotype-aware and target-aware drug design.

Primary area:
generative models
Keywords:
Generative Models, Diffusion Process, Representation learning, Cross-Modal Alignment, Molecular Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_eWNJOpenReview ↗
Rating
3(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors have proposed a representation alignment and fusion framework for molecular generation with diffusion.

Critical strengths
  1. The clear illustration in Figure 1 demonstrated the ICLR-related contributions in methodology.
  2. Molecular generation is a well-sought topic in 2026.
  3. The proposed Blip2Mol framework is well-motivated by its biological relevance.
  4. Source code is opened for reproducibility.
  5. Ablation analysis has been provided.
Critical weaknesses

The main results from Figure 2 to Figure 4 are not very impressive but could be honest findings.

Questions for authors
  1. Statistical significance testing could be considered.
  2. Did you release the 100 generated novel molecules in section 3.6 fully ?
Reviewer_SandOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Blip2Mol, a phenotype-based generative model that learns a multimodal representation by aligning embeddings of perturbed expression profiles, binding pockets, and molecular structures. These embeddings condition the latent diffusion process. The alignment is based on BLIP and uses a contrastive objective and a cross-modal matching loss with hard negative mining. The latent diffusion conditioning is implemented with a trainable gate that fuses two modalities. The experiments show that the proposed model can optimize docking scores while maintaining favorable QED and SA scores. The ablation study compares the final model with models conditioned only on the expression profile or only on the binding pocket, revealing a visible tradeoff between docking score and QED.

Critical strengths
  1. The method combines expression perturbations and binding pocket structures as conditions for the generative process. The architecture is reasonable and enables mixing of both signals with a trainable gate. The unique part of the approach is the BLIP-style cross-modal alignment.
  2. The paper is very clear and easy to follow. The schematic in Figure 1 very clearly describes the proposed architecture.
Critical weaknesses
  1. The experimental evaluation is practically limited to one benchmark and three simple metrics. Sections 3.3 and 3.4 use the same experimental setup, but compare Blip2Mol against different baselines. Metrics such as QED and SA are very easy objectives to optimize, and binding affinity, or more precisely docking score, estimates only how likely the molecule will bind to the target. The paper presents no metrics showing that the model can optimize phenotypes. When only phenotype conditioning is present (Table 1), QED and SA are slightly better, but we do not see values for the model without any conditioning, which might be comparable. I would suggest considering better evaluation of the phenotype conditioning and separately reporting distributions of typical drug-like properties, such as the number of hydrogen bond donors and acceptors and the number of rotatable bonds. The question is whether phenotype-based conditioning can generate novel compounds that are further from the ZINC data but still retain drug-like features.
  2. The novelty of this method is somewhat limited. The main novelty is the application of the BLIP-style loss function to phenotype-guided molecule generation, but the objective is similar to those used in related studies, such as MIGA [1]. The conditional latent diffusion is not new either. I recommend clearly identifying the genuinely novel components of this approach and conducting ablation studies to show that they are essential for strong performance.

[1] Zheng, Shuangjia, et al. "Cross‐Modal Graph Contrastive Learning with Cellular Images." Advanced Science 11.32 (2024): 2404845.

Questions for authors
  1. What is the docking protocol used in this study? What PDB structures did you use?
  2. The generated compounds seem to have different numbers of donors and acceptors. Does phenotype alignment help with obtaining better interactions in the binding pocket, e.g. hydrogen bonds expected from drugs designed for specific targets?
  3. The CrossDocked2020 dataset is an order of magnitude larger than the expression profile dataset you use. Do you employ any balancing strategies? What do you provide to the model when you have a ligand-protein complex but not an expression profile? Or do you use this data only for alignment and use only molecules with both embeddings to train the diffusion model?
Reviewer_MbKBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a molecular generator conditioned on both transcriptomic profiles and protein pockets. A SMILES VAE defines the molecular latent space. Separate encoders align transcriptomic and pocket representations to that space using contrastive and matching losses. A latent diffusion model is first trained with transcriptomic conditioning and then fine-tuned with a learned gate combining both modalities. Evaluation uses genetic perturbation profiles and structures for ten cancer-related targets, with docking scores and standard molecular generation metrics.

The motivation is sensible: molecules that fit a pocket need not produce a useful cellular response. However, the experiments currently test chemical properties and docking, not whether generated molecules reproduce the intended response. This leaves a substantial gap between the motivating problem and the demonstrated result.

Critical strengths
  1. The scientific question matters. Combining target information with cellular responses could address a real weakness of structure-only design. The distinction between binding compatibility and cellular function is well motivated.
  2. The architecture is understandable and modular. Using a fixed molecular latent space provides a practical interface between heterogeneous datasets. The sequential training scheme could be useful when fully paired multimodal examples are scarce.
  3. The paper includes relevant internal controls. Both single-modality variants and a no-diffusion variant are evaluated. Reporting QED and SA alongside docking is preferable to optimizing docking alone.
  4. There is evidence of useful molecular generation. The reported validity is 91–97% across targets, with high uniqueness among valid outputs. Several comparisons indicate favorable QED and SA. These are useful engineering results, even though they do not establish biological activity.
Critical weaknesses

1. The central phenotype claim is not evaluated

Sections 3.2–3.6 evaluate validity, uniqueness, novelty, QED, SA, and LeDock scores. None measures whether a generated molecule induces the requested transcriptomic response. QED and SA describe chemical properties; they cannot establish phenotype fidelity. Similarly, a favorable docking score does not establish target modulation, its direction, or a therapeutic cellular effect.

There is also a consequential distribution shift: training uses compound-induced profiles in MCF7, whereas testing conditions on gene knockdown or overexpression profiles (Section 3.1, p. 5). A model that maps chemical responses back to molecules is not automatically able to interpret a genetic perturbation as a realizable drug response. Knockdown need not phenocopy partial pharmacological inhibition, and binding to TP53 or SMAD3 does not show that a compound reproduces overexpression.

I would like an independent phenotype evaluation, such as held-out compound retrieval or generation assessed with a separately validated response predictor, together with shuffled-profile controls. Ideally, a small experimental study would test both the desired response and target engagement. The immediate requirement is to test the claimed biological relationship; it is not to require a complete drug development study. Without such evidence, the conclusions should be limited to molecular generation under transcriptomic and pocket conditioning.

2. The paired dataset and generalization protocol are insufficiently specified

Section 3.1 gives counts for the three source datasets but does not report how many multimodal samples remain after InChIKey matching, how many unique compounds and targets they contain, or how they are divided into training, validation, and test sets. The size of this intersection is critical because it supports the final multimodal fine-tuning stage.

Please report compound-level, scaffold-level, and target/pocket-level overlap across all training stages, including VAE pretraining and pocket alignment. CrossDocked poses are not independent unique experimentally measured binding examples; repeated ligands and related structures need to be tracked explicitly. The ten evaluation targets may or may not have been seen during alignment or fine-tuning—the manuscript does not establish this. This is an unresolved evaluation issue, not evidence that leakage occurred.

The transcriptomic processing also needs the number of distinct compounds, dose and exposure-time handling, replicate aggregation, quality filtering, and the gene representation used. These choices determine whether the model learns pharmacology or dataset-specific variation. A single cell line and ten targets provide a limited basis for broad generalization claims. Exact training-set novelty does not resolve any of these issues.

3. The ablations show a trade-off, not convincing multimodal synergy

The claim of a “significant synergistic effect” on p. 8 is not supported by the reported comparisons:

  • In Table 1, the full model has better docking means than the pocket-only model on 2 of 10 targets, ties on 2, and is worse on 6.
  • The full model never exceeds the profile-only model in QED: 0 wins, 3 ties, and 7 losses.
  • On TP53, the profile-only model has better reported means on all three metrics: docking −5.75 versus −5.64, QED 0.74 versus 0.72, and SA 2.89 versus 2.94.

These comparisons do not prove that fusion is useless. They show that its benefit needs a more careful definition. Report a prespecified joint success criterion or Pareto analysis, together with per-target uncertainty, instead of describing a mixed trade-off as synergy.

The paper does not define whether the “±” values in Table 1 are standard deviations, standard errors, or variation across independent runs. Sampling 100 molecules per target does not measure training-seed variation. Please repeat training/sampling with independent seeds and report paired target-level effect sizes and confidence intervals. The significance of the small differences cannot be inferred from the current presentation.

There are also direct reporting inconsistencies. The EGFR discussion on p. 8 assigns −7.05 to the full model and calls it the best, whereas Table 1 assigns −7.01 to the full model and −7.05 to the profile-only model. The AURKB full-model score is −6.52 in the prose but −6.57 in Table 1. These should be reconciled before interpreting the ablation.

4. The comparisons do not isolate the proposed methodological contribution

The full model receives both modalities while the external baselines receive one. Such comparisons are useful for the practical system-level question, but they cannot by themselves distinguish an improved learning method from the benefit of additional information or different pretraining data.

The paper itself cites a closely related phenotype–target joint generation method, XMolRL/ExMolRL, in Section 4, but does not compare against it. Please include an appropriately matched comparison or explain why one is infeasible. A simple concatenation-based conditional generator, a fixed-weight fusion model, and a retrieval-and-docking baseline would also help establish what the learned gate and alignment actually contribute.

The technical components—SMILES VAE, contrastive alignment, hard-negative matching, latent diffusion, and gated fusion—are individually established. A valuable contribution could still come from their integration, but it requires clear evidence that this combination solves a previously unresolved problem. There are no separate ablations of the matching objective, contrastive objective, or learned gate. Moreover, Equation (3) applies binary classification to dot-product scores; the manuscript should explain what “BLIP-style” contributes beyond this relatively simple alignment procedure.

For the no-diffusion comparison, specify exactly how latent samples are produced once diffusion is removed. Otherwise, improved validity may reflect how well samples match the decoder's training distribution rather than a controlled demonstration of diffusion's specific advantages.

5. Docking and the therapeutic case study are overinterpreted

The docking protocol needs receptor identifiers and preparation, pocket definitions, protonation/tautomer handling, conformer generation, search settings, and treatment of failed molecules. It is particularly important to establish that the same redocking and filtering pipeline is applied to SMILES-generating and 3D-generating methods.

The selected examples in Figure 5 compare individual generated compounds with reference compounds after generating 100 candidates per target. Please state how those examples were selected and provide the full score distributions and success fractions. A selected favorable docking pose does not establish better affinity, selectivity, or therapeutic utility.

The designation “DrugBank-approved drugs” in Section 3.6 is also problematic. Inclusion in DrugBank is not itself regulatory approval. For example, published clinical evidence describes odanacatib development as discontinued after an increased stroke risk. Conversely, the statement that MTOR lacks approved reference drugs is incorrect as written: everolimus is an approved mTOR inhibitor. Its FKBP12-dependent mechanism may make it unsuitable for a particular docking setup, but that would be a structural/protocol exclusion, not an absence of approved drugs. Please verify the status and mechanism of every reference compound and revise the comparison accordingly.

These factual issues do not invalidate all generation results, but they weaken the clinical interpretation of the case study.

6. Alignment may confuse valid many-to-many associations with negatives

Equations (2)–(3) treat the paired molecule as positive and other batch molecules as negatives, with the most similar mismatch used as a hard negative. However, a pocket can bind multiple molecules, and multiple compounds can induce related expression profiles. The same compound can also appear under several perturbation conditions or in several complexes.

Please describe duplicate handling, batch construction, and whether known related/positive pairs are excluded from negative mining. This is especially relevant after pooling heterogeneous CrossDocked and LINCS records. I am not claiming that every off-diagonal pair is positive; the concern is that the current objective and data description do not address a predictable source of false negatives.

Questions for authors
  1. How many unique compound–profile–pocket triples remain after matching, and which evaluation targets, ligands, and close analogues are present in each training stage?
  2. Does the profile channel improve an independent phenotype endpoint, and does the improvement disappear when profiles are shuffled or replaced with unrelated genetic perturbations?
  3. What do the uncertainty values represent, and which results remain stable across independently trained models?
  4. Can a matched dual-modality baseline reproduce the reported trade-off? What additional benefit comes from the gate and the two alignment losses?
  5. What receptor structures and ligand-context features are used at inference? Section 2.3 mentions “ligand context”; please distinguish permissible pocket-definition information from features requiring a reference ligand.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.